OmniCast Achieves 20× Speed Boost and Eliminates Autoregressive Error Accumulation in S2S Weather Forecasting

Subseasonal‑to‑seasonal (S2S) weather forecasting bridges short‑term weather prediction and long‑term climate projection, targeting 2‑to‑6‑week horizons that are critical for agriculture and disaster mitigation. Traditional numerical weather prediction (NWP) relies on solving complex physical equations, incurring high computational cost, while data‑driven autoregressive models suffer from error accumulation over longer lead times and ignore slow‑varying boundary forcings.

To address these challenges, a team from the University of California, Los Angeles (UCLA) and Argonne National Laboratory introduced OmniCast, a masked latent diffusion model for high‑precision probabilistic S2S forecasts. OmniCast integrates a variational auto‑encoder (VAE) with a Transformer, employing a cross‑temporal joint sampling strategy that mitigates autoregressive error buildup and learns dynamics beyond initial conditions.

Dataset and Pre‑processing

Training and evaluation use the widely adopted ERA5 reanalysis dataset. From ERA5, 69 meteorological variables are extracted, divided into ground variables (2 m temperature, 10 m wind components, mean sea‑level pressure) and atmospheric variables (potential height, temperature, wind components, specific humidity across 13 pressure levels). For the medium‑range task, WeatherBench2 (WB2) provides training data from 1979‑2018, validation in 2019, and testing in 2020 at 0.25° resolution. For the S2S task, ChaosBench supplies training data from 1979‑2020, validation in 2021, and testing in 2022 at 1.40625° resolution.

Model Architecture

OmniCast follows a two‑stage design. The first stage uses a UNet‑based VAE to compress the high‑dimensional input (69 × 128 × 256 for S2S) into low‑dimensional continuous latent tokens (1024 × 8 × 16), achieving a spatial compression ratio of 16. Continuous VAE is chosen over discrete VAE to avoid excessive compression loss for the many weather variables.

The second stage employs a masked generative Transformer with a diffusion head. It adopts a masked autoencoder (MAE) encoder‑decoder architecture, enabling "no‑error‑accumulation" generation by directly modeling the full sequence of latent tokens. The Transformer consists of 16 layers, each with 16 attention heads, hidden dimension 1024, and dropout 0.1. A small MLP diffusion head predicts the distribution of masked latent tokens.

For short‑term forecasts, an auxiliary mean‑squared‑error loss is added on the first 10 latent frames, with an exponentially decaying weight to emphasize early‑time accuracy.

Experimental Evaluation

OmniCast is compared against two groups of baselines: (1) state‑of‑the‑art deep learning models—PanguWeather (PW), GraphCast (GC), Gencast, ClimaX, Stormer; and (2) traditional ensemble NWP systems—UKMO‑ENS, NCEP‑ENS, CMA‑ENS, ECMWF‑ENS, and IFS‑ENS. Metrics include RMSE, ABS BIAS, SSIM for accuracy; physical consistency scores; and probabilistic scores CRPS and SSR.

In short‑lead times, OmniCast’s RMSE and SSIM are slightly lower than the best deep learning baselines, reflecting its training focus on longer horizons. After 10 days, its relative performance improves, reaching parity with ECMWF‑ENS at 10‑day lead time. Bias is consistently near zero across all variables, and physical consistency surpasses other deep learning methods and often exceeds all baselines.

Probabilistic metrics show OmniCast lagging behind ECMWF‑ENS in the first 15 days but overtaking it thereafter. Overall, OmniCast and ECMWF‑ENS emerge as the top performers across accuracy, physical consistency, and probabilistic measures.

Efficiency experiments reveal that OmniCast trains in 4 days on 32 × NVIDIA A100 GPUs, whereas Gencast requires 5 days on 32 × TPUv5e and NeuralGCM needs 10 days on 128 × TPUv5e. In inference, OmniCast completes a 0.25° forecast in 29 seconds (vs. 480 seconds for Gencast) and a 1.0° forecast in 11 seconds (vs. 224 seconds), delivering a 10‑20× speed advantage.

Broader Context

The paper "OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales" was selected for NeurIPS 2025, underscoring its relevance to the AI‑weather research community. The article also references related advances such as the FuXi‑S2S model and teleconnection‑aware LSTM variants, highlighting a rapidly evolving landscape where deep learning increasingly complements traditional NWP.

In summary, OmniCast introduces a new paradigm for S2S forecasting by jointly modeling spatial‑temporal dynamics in a latent space, eliminating autoregressive error accumulation, and achieving substantial computational gains without sacrificing forecast quality.